Claude Code Auto Mode launched today, March 25, 2026. It replaces the previous step-by-step manual approval flow with a built-in classifier model that reviews each operation before execution and intercepts four defined risk categories automatically. For developers running long agentic sessions, this removes the most common friction point in Claude Code workflows without sacrificing control.
What Changed in Claude Code Auto Mode
Before Auto Mode, Claude Code required the developer to approve each consequential action — file writes, shell commands, package installs, git operations. That model made sense for early agentic tooling where trust boundaries were unclear, but it created a bottleneck: a 20-step refactor required 20 interruptions.
Auto Mode introduces a classifier model that runs inline, reviewing each planned operation before Claude Code executes it. Operations that fall outside the four intercepted risk categories proceed without prompting. Operations that match a risk category are paused and surfaced for human review.
The result: routine operations in a well-scoped task run unattended. High-risk operations still require explicit approval.
The Four Risk Categories
Anthropic has not published the full technical specification of the classifier, but the four categories Auto Mode intercepts automatically are:
| Risk Category | Example Operations |
|---|---|
| Destructive file operations | rm -rf, overwriting files outside the project root |
| Network egress | Outbound HTTP requests, webhook calls, data uploads |
| Credential and secret access | Reading .env files, accessing keychain, SSH key usage |
| Privilege escalation | sudo commands, modifying system files, changing permissions |
Operations in these categories pause execution and present the specific command to the developer for approval before proceeding. Everything else — file edits within the repo, test runs, dependency installs scoped to the project, git commits — executes without interruption.
How to Enable Auto Mode
Auto Mode is available today for all Claude Code users. It is not enabled by default — the existing manual approval flow remains the default to preserve backward compatibility for teams with CI/CD integrations or compliance requirements that depend on explicit approval logs.
Enable via CLI flag
claude --auto
This enables Auto Mode for the current session only. All intercepted operations are logged to ~/.claude/auto-mode.log with timestamps and the classifier's risk category assignment.
Enable as persistent default
claude config set autoMode true
This writes the setting to ~/.claude/config.json and applies Auto Mode to all subsequent sessions. To disable:
claude config set autoMode false
Verify current config
claude config get autoMode
Per-project override
For teams that want Auto Mode on by default but need manual approval in specific repos — production infrastructure, secrets management tooling — add a .claude/config.json at the project root:
{
\"autoMode\": false,
\"autoModeOverride\": \"manual\"
}
Project-level config takes precedence over user-level config. This is useful for monorepos where some subdirectories warrant different trust levels.
What Auto Mode Does Not Change
Auto Mode does not change Claude Code's underlying model, context window, or tool access. Claude Code on Opus 4.6 still uses a 1 million token context window and the same tool suite — bash execution, file read/write, web search, and the computer use interface where available.
It also does not change billing. Token consumption is identical whether Auto Mode is on or off — the classifier runs server-side and is not billed as additional tokens to the user.
Compatibility with existing Claude Code workflows is preserved. Shell scripts that invoke claude programmatically, IDE extensions that wrap Claude Code, and CI pipelines using claude --print are all unaffected unless you explicitly pass --auto.
Why This Matters at Claude Code's Scale
Claude Code accounts for 8% of worldwide GitHub commits as of March 2026 — a figure Anthropic cited in the release announcement. At that commit volume, the aggregate developer time lost to manual approval prompts across all sessions is substantial.
The approval bottleneck also had a subtler cost: developers working on long-horizon tasks would lose context between approvals, particularly in sessions approaching the 1M token limit where summarization was already compressing earlier state. Auto Mode keeps execution continuous, which matters most for the multi-hour refactoring and debugging sessions where Claude Code provides the most value.
Audit and Compliance Considerations
For teams in regulated environments, the auto-mode log at ~/.claude/auto-mode.log provides a machine-readable record of every operation the classifier reviewed, the assigned risk category, and the execution decision. Format is newline-delimited JSON:
{\"ts\":\"2026-03-25T09:14:32Z\",\"op\":\"write\",\"path\":\"/project/src/auth.ts\",\"risk\":null,\"action\":\"executed\"}
{\"ts\":\"2026-03-25T09:14:35Z\",\"op\":\"read\",\"path\":\"/project/.env\",\"risk\":\"credential-access\",\"action\":\"paused\"}
Teams that need approval records for audit purposes can keep Auto Mode enabled while using the log as their paper trail, rather than relying on interactive prompts as the approval mechanism.
Auto Mode vs Manual Mode: When to Use Each
| Scenario | Recommended Mode |
|---|---|
| Local development, personal projects | Auto |
| Team repos, standard feature work | Auto |
| Infrastructure-as-code, Terraform, Kubernetes configs | Manual |
| Production secrets or credential rotation | Manual |
| CI/CD pipeline with compliance audit requirements | Manual (or Auto + log export) |
| Exploratory refactoring across large codebase | Auto |
FAQ
Does Auto Mode work with all Claude Code subscription tiers?
Yes. Auto Mode is available on all tiers — the $20/month Pro plan, the $100/month team plan, and the $200/month enterprise plan. The classifier runs identically across tiers.
Can I customize which risk categories trigger a pause?
Not in the initial release. Anthropic has indicated that configurable risk profiles are on the roadmap, but as of March 25, 2026, the four categories are fixed. Per-project autoMode: false overrides are the current workaround for high-sensitivity repos.
Does the classifier ever make mistakes?
Yes. Any classifier has false positives and false negatives. Anthropic recommends reviewing the auto-mode log periodically and filing feedback via claude feedback when the classifier incorrectly categorizes an operation. The classifier model is updated server-side without requiring a Claude Code version bump.
Is Auto Mode available in the VS Code and JetBrains extensions?
Auto Mode is controlled at the Claude Code CLI level. IDE extensions that wrap the CLI inherit the setting from your user or project config. Check your extension version — extensions shipping before March 25, 2026 may need an update to surface Auto Mode status in the UI.
What happens if Claude Code loses connectivity mid-session in Auto Mode?
Claude Code pauses execution at the next operation boundary and surfaces the connectivity error. It does not attempt to retry operations autonomously. Resume the session manually after connectivity is restored.
Next step: Run claude config set autoMode true in your terminal today, then open the auto-mode log after your next session with cat ~/.claude/auto-mode.log | jq to verify which operations the classifier is intercepting.